RBPNet

A CLIP-seq experiment reports, for every nucleotide in the transcriptome, how many times reverse transcription truncated there — a quantitative footprint of where a protein was crosslinked to RNA. Most sequence models throw nearly all of that away: a peak caller collapses the counts into a few thousand intervals, those intervals become bound/unbound labels on windows hundreds of nucleotides wide, and a classifier learns the labels. The contact spans a few bases, so the answer comes back at the wrong resolution, carrying every choice the peak caller made.

RBPNet skips the labels. It is a dilated convolutional network that maps an RNA sequence directly to the distribution of crosslink counts along it, trained by treating the observed per-position counts as a draw from a multinomial parameterised by the predicted probability vector. Because CLIP libraries carry strong technical bias — uridine photoreactivity, contamination by other proteins — the output head predicts two tracks rather than one: a control track fit against the paired size-matched input experiment, and an unobserved target track carrying protein-specific signal. The observed total is their additive mixture, weighted by a coefficient the network also predicts from sequence. Interpreting the target track alone therefore means interpreting the debiased component.

RBPNet was developed by Marc Horlacher and colleagues in Annalisa Marsico's group at Helmholtz Munich, with the Technical University of Munich, the University of Copenhagen, Slovenia's National Institute of Chemistry and the Francis Crick Institute, and published in Genome Biology in 2023. It carries BPNet's sequence-to-signal idea from ChIP-nexus and DNA over to CLIP-seq and RNA, and predates the transformer-based per-protein binding models the catalog holds in TransRBP, RBPSignal and Reformer. Its own successor from the same lab, PARNET, replaces the per-experiment models with a single multi-task network.

#Key Features

• Single-nucleotide output: the model emits one probability per input base, not a score per window, so predictions compare directly against peak-caller calls.
• Explicit bias correction: modelling the size-matched input control as a separate additive component lets the protein-specific signal be read off independently of background.
• No peak caller in the training path: candidate regions come from a lenient Poisson enrichment test over 100 nt windows, which keeps far more of the experiment than peak calls do.
• Arbitrary-length inference: same padding and no pooling anywhere in the network mean a model trained on 300 nt windows runs unchanged over entire genes hundreds of kilobases long.
• Interpretation built in: rbpnet explain returns integrated-gradients attributions per track, and rbpnet impact scores variants from reference-versus-alternate profiles.

#Technical Details

The body takes a one-hot 300 nt sequence into a 1D convolution with 128 filters of width 12, followed by nine residual blocks, each a dilated convolution of 128 filters of width 6 with exponentially increasing dilation, batch normalisation, ReLU and 0.25 dropout. Each output head is a transposed convolution with a single filter of width 20 producing a 300-dimensional probability vector; the loss sums the negative log-likelihoods of the observed total and control counts. Models were trained with Adam at a 0.004 initial learning rate for at most 50 epochs with early stopping, one per experiment, across 103 HepG2 eCLIP datasets from ENCODE — an average of 302,752 candidate windows per dataset, from 7,937 for LARP7 to 1,105,807 for HNRNPC — plus miCLIP in HEK293 and mESC and iCLIP for TDP-43 and PTBP1. Against PureCLIP crosslink sites on held-out chromosomes, RBPNet averaged 0.89 auROC versus 0.74 for the DeepRiPe classifier, and 0.086 average precision versus 0.012. On miCLIP it reached 0.89 and 0.88 auROC for HEK293 and mESC.

#Applications

The practical draw is imputation: a CLIP experiment only sees transcripts expressed in the assayed cell line, and a trained RBPNet model fills in binding profiles for the rest, including viral RNA the assay never encountered. Attribution maps over ENCODE narrow peaks recover in vitro k-mers at recall comparable to the PEKA motif finder, making motif discovery a by-product rather than a separate tool. For variant interpretation, the authors scored 1,570,195 gnomAD SNVs with models for 15 representative proteins, and separately, over 232 MutSpliceDB splicing mutations, found impact scores elevated against local controls for 22 of 40 splicing-related proteins, 15 of them spliceosomal.

#Impact

RBPNet was the first model to predict the nucleotide-wise distribution of CLIP-seq signal as a function of sequence, and the framing it introduced — fit the raw counts, model the control explicitly, interpret the debiased track — is the one its successors build on. The code is MIT licensed, trained models for the ENCODE panel are deposited on Zenodo, and a BioLib web server runs them without installation. The limits are equally clear: each released model is specific to one protein in one cell line, the input is sequence only with no RNA structure channel, and DeepRiPe outperformed the KL-divergence impact score at separating allele-specific-binding variants.